Abstract Informed Multi�Process Prefetching and Caching

Informed prefetching and caching based on application dis� closure of future I�O accesses �hints � can dramatically re� duce the execution time of I�O�intensive applications. A recent study showed that � in the context of a single hint� ing application � prefetching and caching algorithms should adapt to the dynamic load on the disks to obtain the best performance. In this paper � we show how to incorporate adaptivity to disk load into the TIP2 system � which uses cost�bene�t analysis to allocate global resources among mul� tiple processes. We compare the resulting system � which we call TIPTOE �TIP with Temporal Overload Estimators � to Cao et al�s LRU�SP allocation scheme � also modi�ed to in� clude adaptive prefetching. Using disk�accurate trace�driven simulation we show that � averaged over eleven experiments involving pairs of hinting applications � and with data striped over one to ten disks � TIPTOE delivers 7 � lower execution time than LRU�SP. Where the computation and I�O de� mands of each experiment are closely matched � in a two�disk array � TIPTOE delivers 18 � lower execution time.